Improving retinal screenings of patients by neural networks training

The aim of the project is to develop a new service that enables a multifold increase in preventive and control screenings of the ocular fundus (retina) in order to detect diseases or progression as early as possible and thus reduce the subsequent costs of further treatment of patients, which will then be certified as a medical device.

The subject of the project is therefore to develop and validate an expert system using neural network modelling and training to detect pathologies related to diabetic retinopathy over images taken by a fundus camera. The neural networks are trained to classify and segment fundus camera images, i.e., to categorize the images to determine the degree of severity of the disease, including detailed segmentation of specific defects associated with each degree. Thousands of high-resolution images are used for training and validation with appropriate metadata about the defects in the images. The system will then be able to automatically evaluate pathologies in the eye images in a similar quality to that of a fully qualified medical professional (doctor). This expert system capability will then be clinically validated on real data.

Due to the complexity of the problem, the rapidly increasing number of annotated images and the need to robustly evaluate the results, enormous computing power is required. The project uses specialized systems accelerated for AI computations such as the Barbora supercomputer or NVIDIA DGX-2, which are located at IT4Innovations National Supercomputing Center. The planned output of the project is a series of classification and segmentation models that can be integrated into a broader solution being developed by Bonmedix. A major benefit of the project is a massive increase in screening capacity, allowing more patients to be examined in the same time period, without increasing staff costs.

"It now takes about 15 minutes of a specialist retinal ophthalmologist's time to subjectively evaluate  patient's images, with only baseline quantification on top of that. Our solution allows the images to be evaluated in a matter of seconds. The direct cost of such an evaluation with very precise quantification, both numerical and dimensional, timeline and trends, is incomparably lower, as it leads to a substantial saving of the doctor's time - 15 minutes per patient vs. 3 seconds per patient," says Jiří Tomšej from Bonmedix s.r.o.

The project Development of an expert system for automatic evaluation of pathologies from eye images, number FW2020151, is co-financed with state support from the Technology Agency of the Czech Republic under the TREND programme.